--- license: apache-2.0 tags: - robotics - vla - vision-language-action - libero - model-compression pipeline_tag: robotics --- # VLADrop-pi05-LIBERO-vision-drop-half Checkpoint for [Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?](https://arxiv.org/abs/2606.27755). DTR (Drop-Then-Recovery) removes transformer blocks from a pretrained VLA model and recovery-fine-tunes the smaller dense model. Code: https://github.com/s1ghhh/VLADrop ## This checkpoint | | | |---|---| | Paper row | Table 1: pi0.5 Drop Half Vision | | Dropped blocks | Vision encoder (SigLIP, 27 layers): drop all odd-indexed blocks [1,3,5,...,25] (attn+mlp). Language and action untouched. | | Recovery training | batch size 32, 30K steps, lr 5e-5 | | LIBERO success rate | Spatial 84.0 / Object 90.2 / Goal 82.0 / Long 66.8 / Avg 80.8 | ## Usage This is an [openpi](https://github.com/Physical-Intelligence/openpi)-format pi0.5 checkpoint (PyTorch). Use with the VLADrop fork: https://github.com/s1ghhh/VLADrop ```bash python scripts/serve_policy_batch_drop.py \ --config pi05_libero_dropped \ --dir \ --port 8000 ``` **Important:** the drop lists are NOT stored inside the checkpoint. Pass the exact `llm_drop_attn_list` / `llm_drop_mlp_list` shown above (via config or CLI) when serving, otherwise layers will be mismatched. `assets/` contains the LIBERO norm stats. The optimizer state (`train_state/`) is not included. ## Citation ```bibtex @article{sun2026vladrop, title={Drop-Then-Recovery: How Redundant Are Vision-Language-Action Models?}, author={Sun, Guoheng and Feng, Kaixi and He, Shwai and Gong, Xiaochuan and He, Yexiao and Wang, Ziyao and Shen, Zheyu and Ye, Wanghao and Kompella, Ramana Rao and Liu, Gaowen and Li, Ang}, journal={arXiv preprint arXiv:2606.27755}, year={2026} } ```
